MCP Server — Agentic RAG System

A full-stack application that lets users upload documents, chat with an AI agent that retrieves relevant information from those documents, and manage data categories.

---

What It Does

Feature	Description
Upload Files	Upload .txt documents and assign them to a category (cars or addresses)
Chat	Ask questions — the AI agent retrieves relevant chunks from your uploaded documents and answers using GPT-4.1
Remove Category	Delete all vector data stored under a specific namespace/category

---

Architecture Overview

┌──────────────────────────────────────────────────┐
│              Frontend  (React 19 + Vite)          │
│                                                    │
│  ┌────────────┐  ┌──────────┐  ┌───────────────┐  │
│  │ Upload     │  │  Chat    │  │ Remove        │  │
│  │ File       │  │  Page    │  │ Category      │  │
│  └────────────┘  └──────────┘  └───────────────┘  │
└───────────────────────┬──────────────────────────┘
                        │ HTTP / REST
                        ▼
┌──────────────────────────────────────────────────┐
│         FastAPI Backend  (port 8000)              │
│                                                    │
│  POST   /api/v1/upload                            │
│  DELETE /api/v1/delete-namespace                  │
│  POST   /api/v1/agent-response                    │
│                                                    │
│  ┌────────────────────────────────────────────┐   │
│  │  LangGraph Agent  (GPT-4.1 via OpenAI)     │   │
│  │  • Loads conversation history (Postgres)   │   │
│  │  • Connects to MCP Server for tools        │   │
│  │  • ContentFilter middleware (security)     │   │
│  └────────────────────┬───────────────────────┘   │
└───────────────────────┼──────────────────────────┘
          ┌─────────────┼────────────────┐
          ▼             ▼                ▼
   ┌─────────────┐  ┌──────────┐  ┌──────────────┐
   │  MCP Server │  │Postgres  │  │   Celery     │
   │  (port 9000)│  │(memory)  │  │   Worker     │
   │             │  │          │  │              │
   │ Tools:      │  │ Thread   │  │ Async file   │
   │ • namespaces│  │ checkpoint│  │ processing   │
   │ • search_   │  │ storage  │  │ & embedding  │
   │   cars      │  └──────────┘  └──────┬───────┘
   │ • search_   │                        │
   │   addresses │         ┌──────────────┘
   └──────┬──────┘         ▼
          │          ┌──────────┐
          │          │  Redis   │
          │          │ (broker) │
          │          └──────────┘
          ▼
   ┌─────────────┐
   │  Pinecone   │
   │  (vectors)  │
   │             │
   │ Embedding:  │
   │ llama-text- │
   │ embed-v2    │
   └─────────────┘

---

Tech Stack

Backend

Layer	Technology
API Framework	FastAPI
AI Agent	LangChain + LangGraph
LLM	OpenAI GPT-4.1
Vector Store	Pinecone (llama-text-embed-v2 embeddings, cosine similarity)
Conversation Memory	LangGraph Postgres Checkpointer
Async Tasks	Celery + Redis
MCP Server	FastMCP (port 9000)
Database	PostgreSQL 17
Message Broker	Redis 7

Frontend

Layer	Technology
Framework	React 19
Routing	React Router 7
Styling	Tailwind CSS 4
Build Tool	Vite
Language	TypeScript

---

Data Flow — Chat Query

1. User sends a question from the Chat page
2. Frontend POST /api/v1/agent-response?question=...&thread_id=default
3. Backend loads conversation history from PostgreSQL (by thread_id)
4. LangGraph agent initialises with GPT-4.1 and connects to the MCP Server
5. Agent calls retrive_all_name_spaces() to discover available categories
6. Agent decides which category the question is about (cars or addresses)
7. Agent calls search_cars_vectors() or search_addresses_vectors() → Pinecone returns top 3 semantic matches
8. GPT-4.1 generates a final answer using the retrieved context
9. Conversation state saved back to PostgreSQL
10. { "answer": "..." } returned to the frontend

---

Data Flow — File Upload

1. User selects a file and category in the Upload page
2. Frontend POST /api/v1/upload?category=cars with multipart/form-data
3. FastAPI enqueues a Celery task (via Redis broker)
4. Celery worker processes the file:
   • Splits text into ~200-character chunks
   • Embeds each chunk with Pinecone's llama-text-embed-v2
   • Stores vectors under the matching namespace (cars or addresses)
5. Success response returned

---

Project Structure

mcp_server/
├── backend/
│   ├── app/
│   │   ├── main.py              # FastAPI app + CORS + router
│   │   ├── routes.py            # 3 API endpoints
│   │   ├── agent_logic.py       # LangGraph agent, upload logic, namespace removal
│   │   ├── mcp_server.py        # FastMCP server with 3 tools
│   │   ├── celery_tasks.py      # Async Celery tasks for file processing
│   │   ├── pinecone_client.py   # Pinecone connection + embedding
│   │   ├── schemas.py           # Pydantic request/response models
│   │   ├── middleware.py        # ContentFilterMiddleware
│   │   ├── logger.py            # Coloured logging
│   │   └── aws_secretes.py      # AWS Secrets Manager integration
│   ├── docker-compose.yml       # PostgreSQL + Redis + FastAPI + Celery + MCP
│   ├── Dockerfile
│   └── pyproject.toml
├── frontend/
│   ├── src/
│   │   ├── App.tsx              # BrowserRouter + route definitions
│   │   ├── pages/
│   │   │   ├── UploadFilePage.tsx
│   │   │   ├── ChatPage.tsx
│   │   │   └── RemoveCategoryPage.tsx
│   │   └── components/
│   │       ├── Nav.tsx          # Hamburger sidebar navigation
│   │       ├── Layout.tsx
│   │       └── UploadFile.tsx
│   ├── .env                     # VITE_SERVER_URL=http://localhost:8000
│   └── vite.config.ts
└── README.md

---

Running the App

Backend

cd backend
docker compose up --build

This starts 5 services:

• fastapi — API server on port 8000
• mcp_server — MCP tool server on port 9000
• celery_worker — async task processor
• postgres — PostgreSQL 17 on port 5432
• redis — Redis 7 on port 6379

Frontend

cd frontend
npm install
npm run dev

App runs on http://localhost:5173.

Environment Variables

Backend (via Docker or .env):

Variable	Description
OPENAI_API_KEY	OpenAI API key
PINECONE_API_KEY	Pinecone API key
POSTGRES_USER	PostgreSQL username
POSTGRES_PASSWORD	PostgreSQL password
POSTGRES_HOST	PostgreSQL host
POSTGRES_PORT	PostgreSQL port
POSTGRES_DB	PostgreSQL database name

Frontend (.env):

Variable	Description
VITE_SERVER_URL	Backend base URL (default: http://localhost:8000)

---

API Reference

POST /api/v1/upload

Upload a file and embed it into the vector store.

Parameter	Type	Location	Description
user_file	file	form-data	The file to upload
category	cars | addresses	query	Target namespace

Response: { "success": "..." }

---

POST /api/v1/agent-response

Send a question to the AI agent.

Parameter	Type	Location	Description
question	string (max 100)	query	The user's question
thread_id	string	query	Conversation ID for memory (default: "default")

Response: { "answer": "..." }

---

DELETE /api/v1/delete-namespace

Remove all vectors stored under a category.

Parameter	Type	Location	Description
category	string (max 50)	query	Namespace to delete

Response: { "response": "..." }